Following OpenAI's release of Deep Research, a team reproduced its agentic framework in 24 hours and open-sourced the implementation via the smolagents library. The resulting system integrates an LLM with tools for web browsing and text inspection to perform multi-step reasoning tasks.
- The reproduction utilizes a "code agent" architecture, allowing the model to express actions as code rather than JSON, which reduces token usage and improves state handling.
- On the GAIA validation set, the open-source system achieved 55.15%, surpassing the previous state-of-the-art of 46% held by Magentic-One.
- This performance significantly exceeds the ~7% score of standalone GPT-4 on the same benchmark and approaches OpenAI's reported 67.36%.
- The team notes that full parity with OpenAI would require more advanced browser interaction capabilities beyond the current text-only implementation.
The authors consider this significant because it demonstrates that open-source agentic frameworks can dramatically boost LLM performance on complex, knowledge-intensive benchmarks without relying on proprietary models.